... performance¹

This argument has been made since the late 60s, according to [2], p67.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... implicitly²

Depending on whether the platform presents a shared memory or message-passing model.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... FORTRAN³

The discussion here is based on chapter 13 of [8].

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... GFLOPS⁴

In the CM the term ``processing element'' became overloaded. To make best use of the floating point hardware, the software had the ability to optionally treat the 32 PE processing node as a single logical processing element.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... Fortran⁵

With this standard the approved spelling of FORTRAN changed to lower case.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... 90⁶

It did not include all the other new features of Fortran 90.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... FORALL⁷

A similar construct had appeared much earlier in IVTRAN.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... PE⁸

A useful property of the hypercube is that any grid whose extents are powers of two can be embedded in the hypercube graph in such a way that there is an edge between grid neighbours. So a VP grid could always be mapped to the hypercube network so that neighbouring points could communicate directly.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... occurred⁹

Technically, there are no loop-carried flow dependences.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... shapes¹⁰

There is however, an implicit equivalence between processor arrangements of the same shape.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.